Bayesian estimators are defined in terms of the posterior distribution. Typically, this is written as the product of the likelihood function and a prior probability density, both of which are assumed to be known. But in many situations, the prior density is not known, and is difficult to learn from data since one does not have access to uncorrupted samples of the variable being estimated. We show that for a wide variety of observation models, the Bayes least squares (BLS) estimator may be formulated without explicit reference to the prior. Specifically, we derive a direct expression for the estimator, and a related expression for the mean squared estimation error, both in terms of the density of the observed measurements. Each of these prior-free formulations allows us to approximate the estimator given a sufficient amount of observed data. We use the first form to develop practical nonparametric approximations of BLS estimators for several different observation processes, and the second form to develop a parametric family of estimators for use in the additive Gaussian noise case. We examine the empirical performance of these estimators as a function of the amount of observed data. 1

Abstract not found

Bayesian estimators are commonly constructed using an explicit prior model. In many applications, one does not have such a model, and it is difficult to learn since one does not have access to uncorrupted measurements of the variable being estimated. In many cases however, including the case of contamination with additive Gaussian noise, the Bayesian least squares estimator can be formulated directly in terms of the distribution of noisy measurements. We demonstrate the use of this formulation in removing noise from photographic images. We use a local approximation of the noisy measurement distribution by exponentials over adaptively chosen intervals, and derive an estimator from this approximate distribution. We demonstrate through simulations that this adaptive Bayesian estimator performs as well or better than previously published estimators based on simple prior models. 1

Abstract: A variety of experimental studies suggest that sensory systems are capable of performing estimation or decision tasks at near-optimal levels. In this chapter, I explore the use of optimal estimation in describing sensory computations in the brain. I define what is meant by optimality and provide three quite different methods of obtaining an optimal estimator, each based on different assumptions about the nature of the information that is available to constrain the problem. I then discuss how biological systems might go about computing (and learning to compute) optimal estimates. The brain is awash in sensory signals. How does it interpret these signals, so as to extract meaningful and consistent information about the environment? Many tasks require estimation of environmental parameters, and there is substantial evidence that the system is capable of representing and extracting very precise estimates of these parameters. This is particularly impressive when one considers the fact that the brain is built from a large number of low-energy unreliable components, whose responses are affected by many extraneous factors (e.g., temperature, hydration, blood glucose and oxygen levels). The problem of optimal estimation is well studied in the statistics and engineering communities, where a plethora of tools have been developed for designing, implementing, calibrating and testing such systems. In recent years, many of these tools have been used to provide benchmarks or models for biological perception. Specifically, the development of signal detection theory led to widespread use of statistical decision theory as a framework for assessing performance in perceptual experiments. More recently, optimal estimation theory (in particular, Bayesian estimation) has been used as a framework for describing human performance in perceptual tasks.

The two standard methods of obtaining a least-squares optimal estimator are (1) Bayesian estimation, in which one assumes a prior distribution on the true values and combines this with a model of the measurement process to obtain an optimal estimator, and (2) supervised regression, in which one optimizes a parametric estimator over a training set containing pairs of corrupted measurements and their associated true values. But many real-world systems do not have access to either supervised training examples or a prior model. Here, we study the problem of obtaining an optimal estimator given a measurement process with known statistics, and a set of corrupted measurements of random values drawn from an unknown prior. We develop a general form of nonparametric empirical Bayesian estimator that is written as a direct function of the measurement density, with no explicit reference to the prior. We study the observation conditions under which such “prior-free ” estimators may be obtained, and we derive specific forms for a variety of different corruption processes. Each of these prior-free estimators may also be used to express the mean squared estimation error as an expectation over the measurement density, thus generalizing Stein’s unbiased risk estimator (SURE) which provides such an expression for the additive Gaussian noise case. Minimizing this expression over measurement samples provides an “unsupervised

A number of recent algorithms in signal and image processing are based on the empirical distribution of localized patches. Here, we develop a nonparametric empirical Bayesian estimator for recovering an image corrupted by additive Gaussian noise, based on fitting the density over image patches with a local exponential model. The resulting solution is in the form of an adaptively weighted average of the observed patch with the mean of a set of similar patches, and thus both justifies and generalizes the recently proposed nonlocalmeans (NL-means) method for image denoising. Unlike NL-means, our estimator includes a dependency on the size of the patch similarity neighborhood, and we show that this neighborhood size can be chosen in such a way that the estimator converges to the optimal Bayes least squares estimator as the amount of data grows. We demonstrate the increase in performance of our method compared to NL-means on a set of simulated examples. 1